[svn] / branches / dev-api-3 / xvidcore / src / motion / motion_est.c Repository:
ViewVC logotype

Annotation of /branches/dev-api-3/xvidcore/src/motion/motion_est.c

Parent Directory Parent Directory | Revision Log Revision Log


Revision 712 - (view) (download)

1 : Isibaar 3 /**************************************************************************
2 :     *
3 : chl 259 * XVID MPEG-4 VIDEO CODEC
4 :     * motion estimation
5 :     *
6 :     * This program is an implementation of a part of one or more MPEG-4
7 :     * Video tools as specified in ISO/IEC 14496-2 standard. Those intending
8 :     * to use this software module in hardware or software products are
9 :     * advised that its use may infringe existing patents or copyrights, and
10 :     * any such use would be at such party's own risk. The original
11 :     * developer of this software module and his/her company, and subsequent
12 :     * editors and their companies, will have no liability for use of this
13 :     * software or modifications or derivatives thereof.
14 :     *
15 :     * This program is free software; you can redistribute it and/or modify
16 :     * it under the terms of the GNU General Public License as published by
17 :     * the Free Software Foundation; either version 2 of the License, or
18 :     * (at your option) any later version.
19 :     *
20 :     * This program is distributed in the hope that it will be useful,
21 :     * but WITHOUT ANY WARRANTY; without even the implied warranty of
22 :     * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
23 :     * GNU General Public License for more details.
24 :     *
25 :     * You should have received a copy of the GNU General Public License
26 :     * along with this program; if not, write to the Free Software
27 :     * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
28 :     *
29 :     *************************************************************************/
30 :    
31 : Isibaar 3 #include <assert.h>
32 :     #include <stdio.h>
33 : chl 96 #include <stdlib.h>
34 : Isibaar 3
35 :     #include "../encoder.h"
36 :     #include "../utils/mbfunctions.h"
37 :     #include "../prediction/mbprediction.h"
38 :     #include "../global.h"
39 :     #include "../utils/timer.h"
40 : Isibaar 580 #include "../image/interpolate8x8.h"
41 : chl 530 #include "motion_est.h"
42 : suxen_drol 118 #include "motion.h"
43 : Isibaar 3 #include "sad.h"
44 : Isibaar 539 #include "../utils/emms.h"
45 : Isibaar 3
46 : chl 530 #define INITIAL_SKIP_THRESH (10)
47 :     #define FINAL_SKIP_THRESH (50)
48 :     #define MAX_SAD00_FOR_SKIP (20)
49 : Isibaar 539 #define MAX_CHROMA_SAD_FOR_SKIP (22)
50 : Isibaar 3
51 : chl 530 #define CHECK_CANDIDATE(X,Y,D) { \
52 :     (*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); }
53 : Isibaar 3
54 : chl 530 static __inline int
55 : syskin 712 d_mv_bits(int x, int y, const uint32_t iFcode, const int qpel, const int rrv)
56 : Isibaar 3 {
57 : chl 530 int xb, yb;
58 : syskin 712 if (qpel) { x *= 2; y *= 2;}
59 :     else if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); }
60 : chl 530
61 :     if (x == 0) xb = 1;
62 :     else {
63 :     if (x < 0) x = -x;
64 :     x += (1 << (iFcode - 1)) - 1;
65 :     x >>= (iFcode - 1);
66 :     if (x > 32) x = 32;
67 :     xb = mvtab[x] + iFcode;
68 : edgomez 195 }
69 : Isibaar 3
70 : chl 530 if (y == 0) yb = 1;
71 :     else {
72 :     if (y < 0) y = -y;
73 :     y += (1 << (iFcode - 1)) - 1;
74 :     y >>= (iFcode - 1);
75 :     if (y > 32) y = 32;
76 :     yb = mvtab[y] + iFcode;
77 :     }
78 :     return xb + yb;
79 : Isibaar 3 }
80 :    
81 : syskin 628 static int32_t
82 :     ChromaSAD(int dx, int dy, const SearchData * const data)
83 : Isibaar 3 {
84 : syskin 628 int sad;
85 :     dx = (dx >> 1) + roundtab_79[dx & 0x3];
86 :     dy = (dy >> 1) + roundtab_79[dy & 0x3];
87 : Isibaar 3
88 : syskin 712 if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently
89 :    
90 :     switch (((dx & 1) << 1) | (dy & 1)) {
91 : syskin 628 case 0:
92 :     sad = sad8(data->CurU, data->RefCU + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2);
93 :     sad += sad8(data->CurV, data->RefCV + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2);
94 :     break;
95 :     case 1:
96 :     dx = dx / 2; dy = (dy - 1) / 2;
97 :     sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2);
98 :     sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2);
99 :     break;
100 :     case 2:
101 :     dx = (dx - 1) / 2; dy = dy / 2;
102 :     sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2);
103 :     sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2);
104 :     break;
105 :     default:
106 :     dx = (dx - 1) / 2; dy = (dy - 1) / 2;
107 :     interpolate8x8_halfpel_hv(data->RefQ,
108 :     data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2,
109 :     data->rounding);
110 :     sad = sad8(data->CurU, data->RefQ, data->iEdgedWidth/2);
111 :     interpolate8x8_halfpel_hv(data->RefQ,
112 :     data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2,
113 :     data->rounding);
114 :     sad += sad8(data->CurV, data->RefQ, data->iEdgedWidth/2);
115 :     break;
116 : chl 530 }
117 : syskin 712 data->temp[5] = dx; data->temp[6] = dy; data->temp[7] = sad; //backup
118 : syskin 628 return sad;
119 :     }
120 : Isibaar 3
121 : syskin 663 static __inline const uint8_t *
122 :     GetReference(const int x, const int y, const int dir, const SearchData * const data)
123 : Isibaar 606 {
124 : syskin 663 // dir : 0 = forward, 1 = backward
125 :     switch ( (dir << 2) | ((x&1)<<1) | (y&1) ) {
126 :     case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth);
127 :     case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth);
128 :     case 2 : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth);
129 :     case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth);
130 :     case 4 : return data->bRef + x/2 + (y/2)*(data->iEdgedWidth);
131 :     case 5 : return data->bRefV + x/2 + ((y-1)/2)*(data->iEdgedWidth);
132 :     case 6 : return data->bRefH + (x-1)/2 + (y/2)*(data->iEdgedWidth);
133 :     default : return data->bRefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth);
134 : Isibaar 606 }
135 :     }
136 :    
137 : syskin 663 static uint8_t *
138 :     Interpolate8x8qpel(const int x, const int y, const int block, const int dir, const SearchData * const data)
139 : chl 530 {
140 : syskin 663 // create or find a qpel-precision reference picture; return pointer to it
141 : syskin 668 uint8_t * Reference = (uint8_t *)data->RefQ + 16*dir;
142 : syskin 663 const int32_t iEdgedWidth = data->iEdgedWidth;
143 :     const uint32_t rounding = data->rounding;
144 :     const int halfpel_x = x/2;
145 :     const int halfpel_y = y/2;
146 :     const uint8_t *ref1, *ref2, *ref3, *ref4;
147 : chl 326
148 : syskin 663 ref1 = GetReference(halfpel_x, halfpel_y, dir, data); // this reference is used in all cases
149 :     ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
150 :     switch( ((x&1)<<1) + (y&1) ) {
151 :     case 0: // pure halfpel position
152 :     Reference = (uint8_t *) GetReference(halfpel_x, halfpel_y, dir, data);
153 :     Reference += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
154 :     break;
155 : Isibaar 3
156 : syskin 663 case 1: // x halfpel, y qpel - top or bottom during qpel refinement
157 :     ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data);
158 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
159 : Isibaar 665 interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
160 : syskin 663 break;
161 : Isibaar 3
162 : syskin 663 case 2: // x qpel, y halfpel - left or right during qpel refinement
163 :     ref2 = GetReference(x - halfpel_x, halfpel_y, dir, data);
164 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
165 : Isibaar 665 interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
166 : syskin 663 break;
167 : Isibaar 3
168 : syskin 663 default: // x and y in qpel resolution - the "corners" (top left/right and
169 :     // bottom left/right) during qpel refinement
170 :     ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data);
171 :     ref3 = GetReference(x - halfpel_x, halfpel_y, dir, data);
172 :     ref4 = GetReference(x - halfpel_x, y - halfpel_y, dir, data);
173 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
174 :     ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
175 :     ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
176 :     interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
177 :     break;
178 :     }
179 :     return Reference;
180 : Isibaar 3 }
181 :    
182 : syskin 663 static uint8_t *
183 :     Interpolate16x16qpel(const int x, const int y, const int dir, const SearchData * const data)
184 : Isibaar 580 {
185 : syskin 663 // create or find a qpel-precision reference picture; return pointer to it
186 : syskin 668 uint8_t * Reference = (uint8_t *)data->RefQ + 16*dir;
187 : syskin 663 const int32_t iEdgedWidth = data->iEdgedWidth;
188 :     const uint32_t rounding = data->rounding;
189 :     const int halfpel_x = x/2;
190 :     const int halfpel_y = y/2;
191 : Isibaar 580 const uint8_t *ref1, *ref2, *ref3, *ref4;
192 :    
193 : syskin 663 ref1 = GetReference(halfpel_x, halfpel_y, dir, data); // this reference is used in all cases
194 :     switch( ((x&1)<<1) + (y&1) ) {
195 :     case 0: // pure halfpel position
196 :     return (uint8_t *) GetReference(halfpel_x, halfpel_y, dir, data);
197 : Isibaar 580 case 1: // x halfpel, y qpel - top or bottom during qpel refinement
198 : syskin 663 ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data);
199 : Isibaar 665 interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
200 :     interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
201 :     interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
202 :     interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
203 : Isibaar 580 break;
204 :    
205 :     case 2: // x qpel, y halfpel - left or right during qpel refinement
206 : syskin 663 ref2 = GetReference(x - halfpel_x, halfpel_y, dir, data);
207 : Isibaar 665 interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
208 :     interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
209 :     interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
210 :     interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
211 : Isibaar 580 break;
212 :    
213 :     default: // x and y in qpel resolution - the "corners" (top left/right and
214 :     // bottom left/right) during qpel refinement
215 : syskin 663 ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data);
216 :     ref3 = GetReference(x - halfpel_x, halfpel_y, dir, data);
217 :     ref4 = GetReference(x - halfpel_x, y - halfpel_y, dir, data);
218 : Isibaar 580 interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
219 :     interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding);
220 :     interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding);
221 :     interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding);
222 :     break;
223 :     }
224 : syskin 663 return Reference;
225 :     }
226 :    
227 :     /* CHECK_CANDIATE FUNCTIONS START */
228 :    
229 :     static void
230 :     CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
231 :     {
232 :     int t, xc, yc;
233 :     const uint8_t * Reference;
234 :     VECTOR * current;
235 :    
236 :     if (( x > data->max_dx) || ( x < data->min_dx)
237 :     || ( y > data->max_dy) || (y < data->min_dy)) return;
238 :    
239 :     if (data->qpel_precision) { // x and y are in 1/4 precision
240 :     Reference = Interpolate16x16qpel(x, y, 0, data);
241 :     xc = x/2; yc = y/2; //for chroma sad
242 :     current = data->currentQMV;
243 :     } else {
244 : syskin 712 Reference = GetReference(x, y, 0, data);
245 : syskin 663 current = data->currentMV;
246 :     xc = x; yc = y;
247 :     }
248 : syskin 712 t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode, data->qpel && !data->qpel_precision, 0);
249 : Isibaar 580
250 : syskin 663 data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
251 : Isibaar 580
252 : syskin 628 data->temp[0] += (data->lambda16 * t * data->temp[0])/1000;
253 :     data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100;
254 : Isibaar 580
255 : syskin 663 if (data->chroma) data->temp[0] += ChromaSAD(xc, yc, data);
256 : syskin 628
257 : Isibaar 580 if (data->temp[0] < data->iMinSAD[0]) {
258 :     data->iMinSAD[0] = data->temp[0];
259 : syskin 663 current[0].x = x; current[0].y = y;
260 :     *dir = Direction; }
261 : Isibaar 580
262 :     if (data->temp[1] < data->iMinSAD[1]) {
263 : syskin 663 data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y= y; }
264 : Isibaar 580 if (data->temp[2] < data->iMinSAD[2]) {
265 : syskin 663 data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; }
266 : Isibaar 580 if (data->temp[3] < data->iMinSAD[3]) {
267 : syskin 663 data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; }
268 : Isibaar 580 if (data->temp[4] < data->iMinSAD[4]) {
269 : syskin 663 data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; }
270 :    
271 : Isibaar 580 }
272 :    
273 :     static void
274 : syskin 704 CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
275 :     {
276 :     int t;
277 :     const uint8_t * Reference;
278 :    
279 :     if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || //non-zero integer value
280 :     ( x > data->max_dx) || ( x < data->min_dx)
281 :     || ( y > data->max_dy) || (y < data->min_dy)) return;
282 :    
283 : syskin 712 Reference = GetReference(x, y, 0, data);
284 :     t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode, 0, 1);
285 : syskin 704
286 :     data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
287 :    
288 :     data->temp[0] += (data->lambda16 * t * data->temp[0])/1000;
289 :     data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100;
290 :    
291 :     if (data->temp[0] < data->iMinSAD[0]) {
292 :     data->iMinSAD[0] = data->temp[0];
293 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
294 :     *dir = Direction; }
295 :    
296 :     if (data->temp[1] < data->iMinSAD[1]) {
297 :     data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
298 :     if (data->temp[2] < data->iMinSAD[2]) {
299 :     data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
300 :     if (data->temp[3] < data->iMinSAD[3]) {
301 :     data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
302 :     if (data->temp[4] < data->iMinSAD[4]) {
303 :     data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
304 :     }
305 :    
306 :     static void
307 : syskin 663 CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
308 : syskin 658 {
309 :     int32_t sad;
310 : syskin 663 const uint8_t * Reference;
311 :     int t;
312 :     VECTOR * current;
313 : syskin 658
314 :     if (( x > data->max_dx) || ( x < data->min_dx)
315 :     || ( y > data->max_dy) || (y < data->min_dy)) return;
316 :    
317 : syskin 712 if (data->rrv)
318 : syskin 704 if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) ) return; //non-zero integer value
319 :    
320 : syskin 663 if (data->qpel_precision) { // x and y are in 1/4 precision
321 :     Reference = Interpolate16x16qpel(x, y, 0, data);
322 :     current = data->currentQMV;
323 :     } else {
324 : syskin 712 Reference = GetReference(x, y, 0, data);
325 : syskin 663 current = data->currentMV;
326 : syskin 658 }
327 : syskin 712 t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode,
328 :     data->qpel && !data->qpel_precision && !data->rrv, data->rrv);
329 : syskin 658
330 :     sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096);
331 : syskin 663 sad += (data->lambda16 * t * sad)/1000;
332 : syskin 658
333 : syskin 663 if (sad < *(data->iMinSAD)) {
334 :     *(data->iMinSAD) = sad;
335 :     current->x = x; current->y = y;
336 :     *dir = Direction; }
337 : syskin 658 }
338 :    
339 :     static void
340 : Isibaar 539 CheckCandidate16no4vI(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
341 :     {
342 : syskin 658 // maximum speed - for P/B/I decision
343 : Isibaar 539 int32_t sad;
344 :    
345 :     if (( x > data->max_dx) || ( x < data->min_dx)
346 :     || ( y > data->max_dy) || (y < data->min_dy)) return;
347 :    
348 : syskin 601 sad = sad16(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth),
349 : Isibaar 539 data->iEdgedWidth, 256*4096);
350 :    
351 :     if (sad < *(data->iMinSAD)) {
352 :     *(data->iMinSAD) = sad;
353 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
354 :     *dir = Direction; }
355 :     }
356 :    
357 :    
358 :     static void
359 : chl 530 CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data)
360 :     {
361 :     int32_t sad;
362 : syskin 663 int xb, yb, t;
363 : chl 530 const uint8_t *ReferenceF, *ReferenceB;
364 : syskin 663 VECTOR *current;
365 : Isibaar 3
366 : chl 530 if (( xf > data->max_dx) || ( xf < data->min_dx)
367 :     || ( yf > data->max_dy) || (yf < data->min_dy)) return;
368 : Isibaar 3
369 : syskin 663 if (data->qpel_precision) {
370 :     ReferenceF = Interpolate16x16qpel(xf, yf, 0, data);
371 :     xb = data->currentQMV[1].x; yb = data->currentQMV[1].y;
372 :     current = data->currentQMV;
373 :     ReferenceB = Interpolate16x16qpel(xb, yb, 1, data);
374 :     } else {
375 : syskin 712 ReferenceF = GetReference(xf, yf, 0, data);
376 : syskin 663 xb = data->currentMV[1].x; yb = data->currentMV[1].y;
377 : syskin 712 ReferenceB = GetReference(xb, yb, 1, data);
378 : syskin 663 current = data->currentMV;
379 : chl 530 }
380 : edgomez 195
381 : syskin 712 t = d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode, data->qpel && !data->qpel_precision, 0)
382 :     + d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode, data->qpel && !data->qpel_precision, 0);
383 :    
384 : syskin 628 sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
385 : syskin 663 sad += (data->lambda16 * t * sad)/1000;
386 : syskin 628
387 : chl 530 if (sad < *(data->iMinSAD)) {
388 :     *(data->iMinSAD) = sad;
389 : syskin 663 current->x = xf; current->y = yf;
390 : chl 530 *dir = Direction; }
391 : Isibaar 3 }
392 :    
393 : chl 530 static void
394 :     CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
395 : Isibaar 3 {
396 : syskin 628 int32_t sad = 0;
397 : chl 530 int k;
398 :     const uint8_t *ReferenceF;
399 :     const uint8_t *ReferenceB;
400 :     VECTOR mvs, b_mvs;
401 : Isibaar 3
402 : chl 530 if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return;
403 : Isibaar 3
404 : chl 530 for (k = 0; k < 4; k++) {
405 :     mvs.x = data->directmvF[k].x + x;
406 :     b_mvs.x = ((x == 0) ?
407 :     data->directmvB[k].x
408 :     : mvs.x - data->referencemv[k].x);
409 : Isibaar 3
410 : chl 530 mvs.y = data->directmvF[k].y + y;
411 :     b_mvs.y = ((y == 0) ?
412 :     data->directmvB[k].y
413 :     : mvs.y - data->referencemv[k].y);
414 :    
415 :     if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx )
416 :     || ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy )
417 :     || ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx )
418 :     || ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return;
419 : Isibaar 3
420 : syskin 663 if (!data->qpel) {
421 :     mvs.x *= 2; mvs.y *= 2;
422 :     b_mvs.x *= 2; b_mvs.y *= 2; //we move to qpel precision anyway
423 : chl 530 }
424 : syskin 663 ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data);
425 :     ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data);
426 : chl 530
427 :     sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth),
428 : syskin 663 ReferenceF, ReferenceB,
429 : chl 530 data->iEdgedWidth);
430 :     if (sad > *(data->iMinSAD)) return;
431 : edgomez 78 }
432 : chl 530
433 : syskin 712 sad += (data->lambda16 * d_mv_bits(x, y, 1, 0, 0) * sad)/1000;
434 : syskin 628
435 : chl 530 if (sad < *(data->iMinSAD)) {
436 :     *(data->iMinSAD) = sad;
437 :     data->currentMV->x = x; data->currentMV->y = y;
438 :     *dir = Direction; }
439 : Isibaar 3 }
440 :    
441 : chl 530 static void
442 :     CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
443 : chl 96 {
444 : chl 530 int32_t sad;
445 :     const uint8_t *ReferenceF;
446 :     const uint8_t *ReferenceB;
447 :     VECTOR mvs, b_mvs;
448 : chl 96
449 : syskin 576 if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return;
450 : edgomez 195
451 : chl 530 mvs.x = data->directmvF[0].x + x;
452 :     b_mvs.x = ((x == 0) ?
453 :     data->directmvB[0].x
454 :     : mvs.x - data->referencemv[0].x);
455 : edgomez 195
456 : chl 530 mvs.y = data->directmvF[0].y + y;
457 :     b_mvs.y = ((y == 0) ?
458 :     data->directmvB[0].y
459 :     : mvs.y - data->referencemv[0].y);
460 :    
461 :     if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx )
462 :     || ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy )
463 :     || ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx )
464 :     || ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return;
465 : chl 96
466 : syskin 663 if (!data->qpel) {
467 :     mvs.x *= 2; mvs.y *= 2;
468 :     b_mvs.x *= 2; b_mvs.y *= 2; //we move to qpel precision anyway
469 : syskin 712 }
470 : syskin 663 ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data);
471 :     ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data);
472 : chl 96
473 : syskin 628 sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
474 : syskin 712 sad += (data->lambda16 * d_mv_bits(x, y, 1, 0, 0) * sad)/1000;
475 : chl 96
476 : chl 530 if (sad < *(data->iMinSAD)) {
477 :     *(data->iMinSAD) = sad;
478 :     data->currentMV->x = x; data->currentMV->y = y;
479 :     *dir = Direction; }
480 :     }
481 : chl 96
482 : chl 530 static void
483 :     CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
484 :     {
485 : syskin 628 int32_t sad; int t;
486 : chl 530 const uint8_t * Reference;
487 : edgomez 195
488 : chl 530 if (( x > data->max_dx) || ( x < data->min_dx)
489 :     || ( y > data->max_dy) || (y < data->min_dy)) return;
490 : chl 96
491 : syskin 663 if (data->qpel) Reference = Interpolate16x16qpel(x, y, 0, data);
492 : syskin 712 else Reference = GetReference(x, y, 0, data);
493 : edgomez 195
494 : chl 530 sad = sad8(data->Cur, Reference, data->iEdgedWidth);
495 : syskin 712 t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode, data->qpel && !data->qpel_precision, 0);
496 : edgomez 195
497 : syskin 628 sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))/100;
498 : chl 96
499 : Isibaar 606 if (sad < *(data->iMinSAD)) {
500 :     *(data->iMinSAD) = sad;
501 :     data->currentMV->x = x; data->currentMV->y = y;
502 :     *dir = Direction; }
503 :     }
504 :    
505 : Isibaar 539 /* CHECK_CANDIATE FUNCTIONS END */
506 : chl 96
507 : chl 530 /* MAINSEARCH FUNCTIONS START */
508 : edgomez 195
509 : chl 530 static void
510 :     AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection)
511 : chl 181 {
512 :    
513 :     /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
514 :    
515 : syskin 712 int iDirection;
516 : edgomez 195
517 : syskin 712 for(;;) { //forever
518 :     iDirection = 0;
519 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
520 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
521 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
522 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
523 : chl 181
524 : syskin 712 /* now we're doing diagonal checks near our candidate */
525 : chl 181
526 : syskin 712 if (iDirection) { //checking if anything found
527 :     bDirection = iDirection;
528 :     iDirection = 0;
529 :     x = data->currentMV->x; y = data->currentMV->y;
530 :     if (bDirection & 3) { //our candidate is left or right
531 :     CHECK_CANDIDATE(x, y + iDiamondSize, 8);
532 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
533 :     } else { // what remains here is up or down
534 :     CHECK_CANDIDATE(x + iDiamondSize, y, 2);
535 :     CHECK_CANDIDATE(x - iDiamondSize, y, 1);
536 :     }
537 : chl 181
538 : syskin 712 if (iDirection) {
539 :     bDirection += iDirection;
540 :     x = data->currentMV->x; y = data->currentMV->y;
541 : chl 181 }
542 : syskin 712 } else { //about to quit, eh? not so fast....
543 :     switch (bDirection) {
544 :     case 2:
545 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
546 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
547 :     break;
548 :     case 1:
549 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
550 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
551 :     break;
552 :     case 2 + 4:
553 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
554 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
555 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
556 :     break;
557 :     case 4:
558 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
559 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
560 :     break;
561 :     case 8:
562 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
563 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
564 :     break;
565 :     case 1 + 4:
566 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
567 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
568 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
569 :     break;
570 :     case 2 + 8:
571 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
572 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
573 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
574 :     break;
575 :     case 1 + 8:
576 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
577 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
578 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
579 :     break;
580 :     default: //1+2+4+8 == we didn't find anything at all
581 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
582 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
583 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
584 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
585 :     break;
586 :     }
587 :     if (!iDirection) break; //ok, the end. really
588 :     bDirection = iDirection;
589 :     x = data->currentMV->x; y = data->currentMV->y;
590 : chl 181 }
591 : syskin 712 }
592 : chl 181 }
593 :    
594 : chl 530 static void
595 :     SquareSearch(int x, int y, const SearchData * const data, int bDirection)
596 : chl 326 {
597 : chl 530 int iDirection;
598 : chl 326
599 : chl 530 do {
600 :     iDirection = 0;
601 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64);
602 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128);
603 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32);
604 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128);
605 :     if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64);
606 :     if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128);
607 :     if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128);
608 :     if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128);
609 : chl 326
610 : chl 530 bDirection = iDirection;
611 :     x = data->currentMV->x; y = data->currentMV->y;
612 :     } while (iDirection);
613 : chl 346 }
614 :    
615 : chl 530 static void
616 :     DiamondSearch(int x, int y, const SearchData * const data, int bDirection)
617 : chl 346 {
618 :    
619 : chl 181 /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
620 :    
621 : syskin 712 int iDirection;
622 : edgomez 195
623 : syskin 712 do {
624 :     iDirection = 0;
625 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
626 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
627 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
628 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
629 : chl 181
630 : syskin 712 /* now we're doing diagonal checks near our candidate */
631 : chl 181
632 : syskin 712 if (iDirection) { //checking if anything found
633 :     bDirection = iDirection;
634 :     iDirection = 0;
635 :     x = data->currentMV->x; y = data->currentMV->y;
636 :     if (bDirection & 3) { //our candidate is left or right
637 :     CHECK_CANDIDATE(x, y + iDiamondSize, 8);
638 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
639 :     } else { // what remains here is up or down
640 :     CHECK_CANDIDATE(x + iDiamondSize, y, 2);
641 :     CHECK_CANDIDATE(x - iDiamondSize, y, 1);
642 : chl 181 }
643 : syskin 712 bDirection += iDirection;
644 :     x = data->currentMV->x; y = data->currentMV->y;
645 : chl 181 }
646 : syskin 712 }
647 :     while (iDirection);
648 : chl 181 }
649 :    
650 : chl 530 /* MAINSEARCH FUNCTIONS END */
651 : chl 181
652 : chl 530 /* HALFPELREFINE COULD BE A MAINSEARCH FUNCTION, BUT THERE IS NO NEED FOR IT */
653 :    
654 :     static void
655 : syskin 663 SubpelRefine(const SearchData * const data)
656 : chl 96 {
657 : syskin 663 /* Do a half-pel or q-pel refinement */
658 :     VECTOR backupMV;
659 : chl 530 int iDirection; //not needed
660 : chl 96
661 : syskin 663 if (data->qpel_precision)
662 :     backupMV = *(data->currentQMV);
663 :     else backupMV = *(data->currentMV);
664 : edgomez 195
665 : Isibaar 580 CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0);
666 :     CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0);
667 :     CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0);
668 :     CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0);
669 :    
670 :     CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0);
671 :     CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0);
672 :    
673 :     CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0);
674 :     CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0);
675 :     }
676 :    
677 : chl 530 static __inline int
678 :     SkipDecisionP(const IMAGE * current, const IMAGE * reference,
679 :     const int x, const int y,
680 : syskin 704 const uint32_t iEdgedWidth, const uint32_t iQuant, int rrv)
681 : chl 96
682 : Isibaar 3 {
683 : chl 530 /* keep repeating checks for all b-frames before this P frame,
684 :     to make sure that SKIP is possible (todo)
685 :     how: if skip is not possible set sad00 to a very high value */
686 : syskin 704 if(rrv) {
687 :     uint32_t sadC = sad16(current->u + x*16 + y*(iEdgedWidth/2)*16,
688 :     reference->u + x*16 + y*(iEdgedWidth/2)*16, iEdgedWidth/2, 256*4096);
689 :     if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
690 :     sadC += sad16(current->v + (x + y*(iEdgedWidth/2))*16,
691 :     reference->v + (x + y*(iEdgedWidth/2))*16, iEdgedWidth/2, 256*4096);
692 :     if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
693 :     return 1;
694 :     } else {
695 :     uint32_t sadC = sad8(current->u + x*8 + y*(iEdgedWidth/2)*8,
696 :     reference->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2);
697 :     if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
698 :     sadC += sad8(current->v + (x + y*(iEdgedWidth/2))*8,
699 :     reference->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2);
700 :     if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
701 :     return 1;
702 :     }
703 : chl 530 }
704 : edgomez 195
705 : chl 530 static __inline void
706 :     SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad)
707 :     {
708 :     pMB->mode = MODE_NOT_CODED;
709 : Isibaar 539 pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0;
710 :     pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0;
711 : syskin 628
712 : Isibaar 580 pMB->qmvs[0].x = pMB->qmvs[1].x = pMB->qmvs[2].x = pMB->qmvs[3].x = 0;
713 :     pMB->qmvs[0].y = pMB->qmvs[1].y = pMB->qmvs[2].y = pMB->qmvs[3].y = 0;
714 :    
715 : chl 530 pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
716 : Isibaar 3 }
717 :    
718 : syskin 628 bool
719 : chl 530 MotionEstimation(MBParam * const pParam,
720 :     FRAMEINFO * const current,
721 :     FRAMEINFO * const reference,
722 :     const IMAGE * const pRefH,
723 :     const IMAGE * const pRefV,
724 :     const IMAGE * const pRefHV,
725 :     const uint32_t iLimit)
726 : Isibaar 3 {
727 : chl 530 MACROBLOCK *const pMBs = current->mbs;
728 :     const IMAGE *const pCurrent = &current->image;
729 :     const IMAGE *const pRef = &reference->image;
730 : Isibaar 3
731 : chl 530 const VECTOR zeroMV = { 0, 0 };
732 : Isibaar 3
733 : syskin 704 uint32_t mb_width = pParam->mb_width;
734 :     uint32_t mb_height = pParam->mb_height;
735 : suxen_drol 702
736 : chl 530 uint32_t x, y;
737 :     uint32_t iIntra = 0;
738 : syskin 628 int32_t InterBias, quant = current->quant, sad00;
739 : Isibaar 580 uint8_t *qimage;
740 : edgomez 195
741 : Isibaar 539 // some pre-initialized thingies for SearchP
742 : syskin 712 int32_t temp[8];
743 : Isibaar 539 VECTOR currentMV[5];
744 : Isibaar 580 VECTOR currentQMV[5];
745 : Isibaar 539 int32_t iMinSAD[5];
746 :     SearchData Data;
747 : syskin 704 memset(&Data, 0, sizeof(SearchData));
748 : Isibaar 539 Data.iEdgedWidth = pParam->edged_width;
749 :     Data.currentMV = currentMV;
750 : Isibaar 580 Data.currentQMV = currentQMV;
751 : Isibaar 539 Data.iMinSAD = iMinSAD;
752 :     Data.temp = temp;
753 :     Data.iFcode = current->fcode;
754 : Isibaar 580 Data.rounding = pParam->m_rounding_type;
755 : syskin 628 Data.qpel = pParam->m_quarterpel;
756 :     Data.chroma = current->global_flags & XVID_ME_COLOUR;
757 : syskin 704 Data.rrv = current->global_flags & XVID_REDUCED;
758 : Isibaar 539
759 : syskin 704 if ((current->global_flags & XVID_REDUCED)) {
760 : suxen_drol 702 mb_width = (pParam->width + 31) / 32;
761 :     mb_height = (pParam->height + 31) / 32;
762 : syskin 704 Data.qpel = Data.chroma = 0;
763 : suxen_drol 702 }
764 :    
765 : Isibaar 580 if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL)
766 : syskin 594 return 1; // allocate some mem for qpel interpolated blocks
767 : Isibaar 580 // somehow this is dirty since I think we shouldn't use malloc outside
768 :     // encoder_create() - so please fix me!
769 : syskin 628 Data.RefQ = qimage;
770 : chl 530 if (sadInit) (*sadInit) ();
771 : edgomez 195
772 : suxen_drol 702 for (y = 0; y < mb_height; y++) {
773 :     for (x = 0; x < mb_width; x++) {
774 : chl 530 MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width];
775 : syskin 628
776 : syskin 704 if (Data.rrv) pMB->sad16 =
777 :     sad32v_c(pCurrent->y + (x + y * pParam->edged_width) * 32,
778 :     pRef->y + (x + y * pParam->edged_width) * 32,
779 :     pParam->edged_width, pMB->sad8 );
780 :    
781 :     else pMB->sad16 =
782 :     sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16,
783 : chl 530 pRef->y + (x + y * pParam->edged_width) * 16,
784 :     pParam->edged_width, pMB->sad8 );
785 : edgomez 195
786 : syskin 628 if (Data.chroma) {
787 :     pMB->sad16 += sad8(pCurrent->u + x*8 + y*(pParam->edged_width/2)*8,
788 :     pRef->u + x*8 + y*(pParam->edged_width/2)*8, pParam->edged_width/2);
789 :    
790 :     pMB->sad16 += sad8(pCurrent->v + (x + y*(pParam->edged_width/2))*8,
791 :     pRef->v + (x + y*(pParam->edged_width/2))*8, pParam->edged_width/2);
792 :     }
793 :    
794 :     sad00 = pMB->sad16; //if no gmc; else sad00 = (..)
795 :    
796 : chl 530 if (!(current->global_flags & XVID_LUMIMASKING)) {
797 :     pMB->dquant = NO_CHANGE;
798 : syskin 601 pMB->quant = current->quant;
799 :     } else {
800 : syskin 574 if (pMB->dquant != NO_CHANGE) {
801 :     quant += DQtab[pMB->dquant];
802 :     if (quant > 31) quant = 31;
803 :     else if (quant < 1) quant = 1;
804 :     }
805 : syskin 601 pMB->quant = quant;
806 :     }
807 : chl 181
808 : syskin 628 //initial skip decision
809 : chl 619 /* no early skip for GMC (global vector = skip vector is unknown!) */
810 :     if (current->coding_type == P_VOP) { /* no fast SKIP for S(GMC)-VOPs */
811 : syskin 704 if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH * (Data.rrv ? 4:1) )
812 :     if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant, Data.rrv)) {
813 : chl 619 SkipMacroblockP(pMB, sad00);
814 :     continue;
815 :     }
816 :     }
817 : syskin 628
818 :     SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
819 : chl 530 y, current->motion_flags, pMB->quant,
820 : Isibaar 539 &Data, pParam, pMBs, reference->mbs,
821 : chl 530 current->global_flags & XVID_INTER4V, pMB);
822 : edgomez 195
823 : chl 530 /* final skip decision, a.k.a. "the vector you found, really that good?" */
824 : chl 619 if (current->coding_type == P_VOP) {
825 :     if ( (pMB->dquant == NO_CHANGE) && (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP)
826 : syskin 704 && ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1)) )
827 :     if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant, Data.rrv)) {
828 : chl 619 SkipMacroblockP(pMB, sad00);
829 :     continue;
830 :     }
831 :     }
832 : syskin 628
833 : chl 530 /* finally, intra decision */
834 : Isibaar 3
835 : chl 530 InterBias = MV16_INTER_BIAS;
836 : syskin 639 if (pMB->quant > 8) InterBias += 100 * (pMB->quant - 8); // to make high quants work
837 : chl 530 if (y != 0)
838 : syskin 639 if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80;
839 : chl 530 if (x != 0)
840 : syskin 639 if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80;
841 : syskin 628
842 :     if (Data.chroma) InterBias += 50; // to compensate bigger SAD
843 : syskin 704 if (Data.rrv) InterBias *= 4; //??
844 : Isibaar 3
845 : chl 530 if (InterBias < pMB->sad16) {
846 : syskin 704 int32_t deviation;
847 :     if (Data.rrv) {
848 :     deviation = dev16(pCurrent->y + (x + y * pParam->edged_width) * 32,
849 :     pParam->edged_width)
850 :     + dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16,
851 :     pParam->edged_width)
852 :     + dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16 * pParam->edged_width,
853 :     pParam->edged_width)
854 :     + dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16 * (pParam->edged_width+1),
855 :     pParam->edged_width);
856 :     } else
857 :     deviation = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16,
858 : chl 530 pParam->edged_width);
859 : Isibaar 3
860 : syskin 704 if (deviation < (pMB->sad16 - InterBias)) {
861 : Isibaar 580 if (++iIntra >= iLimit) { free(qimage); return 1; }
862 : chl 530 pMB->mode = MODE_INTRA;
863 : Isibaar 539 pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] =
864 :     pMB->mvs[3] = zeroMV;
865 : Isibaar 580 pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] =
866 :     pMB->qmvs[3] = zeroMV;
867 : chl 530 pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] =
868 :     pMB->sad8[3] = 0;
869 :     }
870 :     }
871 :     }
872 : Isibaar 3 }
873 : Isibaar 580 free(qimage);
874 : chl 619
875 :     if (current->coding_type == S_VOP) /* first GMC step only for S(GMC)-VOPs */
876 : syskin 628 current->GMC_MV = GlobalMotionEst( pMBs, pParam, current->fcode );
877 : chl 619 else
878 :     current->GMC_MV = zeroMV;
879 :    
880 : syskin 628 return 0;
881 : chl 530 }
882 : Isibaar 3
883 : edgomez 195
884 : chl 530 #define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16)
885 : Isibaar 3
886 : chl 530 static __inline int
887 :     make_mask(const VECTOR * const pmv, const int i)
888 :     {
889 : Isibaar 539 int mask = 255, j;
890 : chl 530 for (j = 0; j < i; j++) {
891 :     if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already
892 :     if (pmv[i].x == pmv[j].x) {
893 :     if (pmv[i].y == pmv[j].y + iDiamondSize) { mask &= ~4; continue; }
894 :     if (pmv[i].y == pmv[j].y - iDiamondSize) { mask &= ~8; continue; }
895 :     } else
896 :     if (pmv[i].y == pmv[j].y) {
897 :     if (pmv[i].x == pmv[j].x + iDiamondSize) { mask &= ~1; continue; }
898 :     if (pmv[i].x == pmv[j].x - iDiamondSize) { mask &= ~2; continue; }
899 :     }
900 : Isibaar 3 }
901 : chl 530 return mask;
902 :     }
903 : edgomez 195
904 : chl 530 static __inline void
905 : syskin 704 PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount,
906 :     int iHcount, const MACROBLOCK * const prevMB, int rrv)
907 : chl 530 {
908 : edgomez 195
909 : chl 530 //this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself
910 : syskin 704 if (rrv) { iWcount /= 2; iHcount /= 2; }
911 : edgomez 195
912 : syskin 704 if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour
913 : chl 530 pmv[5].x = EVEN(pmv[3].x);
914 : Isibaar 539 pmv[5].y = EVEN(pmv[3].y);
915 :     } else pmv[5].x = pmv[5].y = 0;
916 : chl 169
917 : chl 530 if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour
918 :     else pmv[3].x = pmv[3].y = 0;
919 : Isibaar 3
920 : chl 530 if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour
921 :     else pmv[4].x = pmv[4].y = 0;
922 : chl 169
923 : chl 530 // [1] median prediction
924 : syskin 704 if (rrv) { //median is in halfzero-precision
925 :     pmv[1].x = RRV_MV_SCALEUP(pmv[0].x);
926 :     pmv[1].y = RRV_MV_SCALEUP(pmv[0].y);
927 :     } else { pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); }
928 : chl 169
929 : chl 530 pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask
930 : chl 169
931 : chl 530 pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame
932 :     pmv[2].y = EVEN(prevMB->mvs[0].y);
933 : chl 169
934 : syskin 704 if ((x < iWcount-1) && (y < iHcount-1)) {
935 : chl 530 pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame
936 : Isibaar 539 pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y);
937 :     } else pmv[6].x = pmv[6].y = 0;
938 : syskin 704
939 :     if (rrv) {
940 :     int i;
941 :     for (i = 0; i < 7; i++) {
942 :     pmv[i].x = RRV_MV_SCALEDOWN(pmv[i].x);
943 :     pmv[i].x = RRV_MV_SCALEUP(pmv[i].x); // a trick
944 :     }
945 :     }
946 : chl 530 }
947 : chl 169
948 : chl 530 static void
949 : syskin 628 SearchP(const IMAGE * const pRef,
950 : chl 530 const uint8_t * const pRefH,
951 :     const uint8_t * const pRefV,
952 :     const uint8_t * const pRefHV,
953 :     const IMAGE * const pCur,
954 :     const int x,
955 :     const int y,
956 :     const uint32_t MotionFlags,
957 :     const uint32_t iQuant,
958 : Isibaar 539 SearchData * const Data,
959 : chl 530 const MBParam * const pParam,
960 :     const MACROBLOCK * const pMBs,
961 :     const MACROBLOCK * const prevMBs,
962 :     int inter4v,
963 :     MACROBLOCK * const pMB)
964 :     {
965 : chl 169
966 : chl 530 int i, iDirection = 255, mask, threshA;
967 : Isibaar 539 VECTOR pmv[7];
968 : Isibaar 3
969 : syskin 712 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
970 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
971 : Isibaar 3
972 : syskin 712 get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); //has to be changed to get_pmv(2)()
973 : syskin 628
974 : syskin 712 Data->temp[5] = Data->temp[7] = 256*4096; // to reset chroma-sad cache
975 : syskin 704 if (Data->rrv) i = 2; else i = 1;
976 :     Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i;
977 :     Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
978 :     Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
979 : Isibaar 3
980 : syskin 704 Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16*i;
981 :     Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16*i;
982 :     Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16*i;
983 :     Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16*i;
984 :     Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
985 :     Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
986 :    
987 : syskin 601 Data->lambda16 = lambda_vec16[iQuant];
988 : Isibaar 606 Data->lambda8 = lambda_vec8[iQuant];
989 : syskin 663 Data->qpel_precision = 0;
990 : chl 169
991 : edgomez 568 if (pMB->dquant != NO_CHANGE) inter4v = 0;
992 : edgomez 195
993 : Isibaar 580 for(i = 0; i < 5; i++)
994 :     Data->currentMV[i].x = Data->currentMV[i].y = 0;
995 : Isibaar 539
996 : syskin 663 if (pParam->m_quarterpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0);
997 :     else Data->predMV = pmv[0];
998 : Isibaar 606
999 : syskin 712 i = d_mv_bits(Data->predMV.x, Data->predMV.y, Data->iFcode, 0, 0);
1000 : syskin 628 Data->iMinSAD[0] = pMB->sad16 + (Data->lambda16 * i * pMB->sad16)/1000;
1001 :     Data->iMinSAD[1] = pMB->sad8[0] + (Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS))/100;
1002 : Isibaar 539 Data->iMinSAD[2] = pMB->sad8[1];
1003 :     Data->iMinSAD[3] = pMB->sad8[2];
1004 :     Data->iMinSAD[4] = pMB->sad8[3];
1005 :    
1006 : chl 530 if ((x == 0) && (y == 0)) threshA = 512;
1007 :     else {
1008 : edgomez 568 threshA = Data->temp[0]; // that's when we keep this SAD atm
1009 : chl 530 if (threshA < 512) threshA = 512;
1010 :     if (threshA > 1024) threshA = 1024; }
1011 : Isibaar 3
1012 : chl 530 PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height,
1013 : syskin 704 prevMBs + x + y * pParam->mb_width, Data->rrv);
1014 : edgomez 195
1015 : syskin 704 if (Data->rrv) CheckCandidate = CheckCandidate32;
1016 :     else if (inter4v || Data->chroma) CheckCandidate = CheckCandidate16;
1017 :     else CheckCandidate = CheckCandidate16no4v; //for extra speed
1018 : edgomez 568
1019 : chl 530 /* main loop. checking all predictions */
1020 :    
1021 :     for (i = 1; i < 7; i++) {
1022 :     if (!(mask = make_mask(pmv, i)) ) continue;
1023 : edgomez 568 (*CheckCandidate)(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1024 : Isibaar 539 if (Data->iMinSAD[0] <= threshA) break;
1025 : chl 530 }
1026 :    
1027 : Isibaar 539 if ((Data->iMinSAD[0] <= threshA) ||
1028 :     (MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) &&
1029 :     (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) {
1030 : chl 530 inter4v = 0;
1031 : Isibaar 539 } else {
1032 : Isibaar 3
1033 : syskin 712 MainSearchFunc * MainSearchPtr;
1034 : Isibaar 539 if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1035 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1036 :     else MainSearchPtr = DiamondSearch;
1037 : chl 181
1038 : Isibaar 539 (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1039 : Isibaar 3
1040 : chl 530 /* extended search, diamond starting in 0,0 and in prediction.
1041 :     note that this search is/might be done in halfpel positions,
1042 :     which makes it more different than the diamond above */
1043 : chl 259
1044 : Isibaar 539 if (MotionFlags & PMV_EXTSEARCH16) {
1045 :     int32_t bSAD;
1046 :     VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0];
1047 : syskin 704 if (Data->rrv) {
1048 :     startMV.x = RRV_MV_SCALEUP(startMV.x);
1049 :     startMV.y = RRV_MV_SCALEUP(startMV.y);
1050 :     } else
1051 :     if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel?
1052 :     startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y);
1053 : Isibaar 539 if (!(MVequal(startMV, backupMV))) {
1054 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1055 : Isibaar 3
1056 : Isibaar 606 (*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data);
1057 : Isibaar 539 (*MainSearchPtr)(startMV.x, startMV.y, Data, 255);
1058 :     if (bSAD < Data->iMinSAD[0]) {
1059 :     Data->currentMV[0] = backupMV;
1060 :     Data->iMinSAD[0] = bSAD; }
1061 :     }
1062 : Isibaar 3
1063 : Isibaar 539 backupMV = Data->currentMV[0];
1064 : syskin 704 if (!MotionFlags & PMV_HALFPELREFINE16 || Data->rrv) startMV.x = startMV.y = 0;
1065 :     else startMV.x = startMV.y = 1;
1066 : Isibaar 539 if (!(MVequal(startMV, backupMV))) {
1067 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1068 : edgomez 195
1069 : Isibaar 606 (*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data);
1070 : Isibaar 539 (*MainSearchPtr)(startMV.x, startMV.y, Data, 255);
1071 :     if (bSAD < Data->iMinSAD[0]) {
1072 :     Data->currentMV[0] = backupMV;
1073 :     Data->iMinSAD[0] = bSAD; }
1074 :     }
1075 : Isibaar 3 }
1076 :     }
1077 :    
1078 : syskin 663 if (MotionFlags & PMV_HALFPELREFINE16) SubpelRefine(Data);
1079 : Isibaar 3
1080 : Isibaar 580 for(i = 0; i < 5; i++) {
1081 :     Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors
1082 :     Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
1083 :     }
1084 :    
1085 : syskin 704 if((!Data->rrv) && (pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) {
1086 : Isibaar 580
1087 : syskin 663 Data->qpel_precision = 1;
1088 : syskin 712 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1089 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
1090 : syskin 601
1091 : syskin 663 SubpelRefine(Data);
1092 : Isibaar 580 }
1093 :    
1094 : syskin 628 if (Data->iMinSAD[0] < (int32_t)iQuant * 30 ) inter4v = 0;
1095 : Isibaar 539 if (inter4v) {
1096 :     SearchData Data8;
1097 :     Data8.iFcode = Data->iFcode;
1098 : Isibaar 606 Data8.lambda8 = Data->lambda8;
1099 : Isibaar 539 Data8.iEdgedWidth = Data->iEdgedWidth;
1100 : Isibaar 606 Data8.RefQ = Data->RefQ;
1101 : syskin 628 Data8.qpel = Data->qpel;
1102 : syskin 704 Data8.rrv = Data->rrv;
1103 : Isibaar 539 Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8);
1104 :     Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8);
1105 :     Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8);
1106 :     Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8);
1107 : syskin 628
1108 :     if (Data->chroma) {
1109 : syskin 663 int sumx, sumy, dx, dy;
1110 : syskin 628
1111 :     if(pParam->m_quarterpel) {
1112 : syskin 663 sumx= pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2;
1113 :     sumy = pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2;
1114 :     } else {
1115 :     sumx = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x;
1116 :     sumy = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y;
1117 :     }
1118 :     dx = (sumx >> 3) + roundtab_76[sumx & 0xf];
1119 :     dy = (sumy >> 3) + roundtab_76[sumy & 0xf];
1120 : syskin 628
1121 :     Data->iMinSAD[1] += ChromaSAD(dx, dy, Data);
1122 :     }
1123 : Isibaar 539 }
1124 : Isibaar 3
1125 : syskin 704 if (Data->rrv) {
1126 :     Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x);
1127 :     Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y);
1128 :     }
1129 : chl 530 if (!(inter4v) ||
1130 : Isibaar 539 (Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] +
1131 :     Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) {
1132 : chl 530 // INTER MODE
1133 :     pMB->mode = MODE_INTER;
1134 : Isibaar 539 pMB->mvs[0] = pMB->mvs[1]
1135 :     = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
1136 : Isibaar 3
1137 : chl 530 pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] =
1138 : Isibaar 539 pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0];
1139 : Isibaar 3
1140 : Isibaar 580 if(pParam->m_quarterpel) {
1141 : syskin 663 pMB->qmvs[0] = pMB->qmvs[1]
1142 :     = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0];
1143 :     pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x;
1144 :     pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y;
1145 : syskin 628 } else {
1146 : Isibaar 580 pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
1147 :     pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
1148 :     }
1149 : edgomez 195 } else {
1150 : chl 530 // INTER4V MODE; all other things are already set in Search8
1151 :     pMB->mode = MODE_INTER4V;
1152 : Isibaar 539 pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] +
1153 :     Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * iQuant;
1154 : edgomez 78 }
1155 : Isibaar 3 }
1156 :    
1157 : chl 530 static void
1158 :     Search8(const SearchData * const OldData,
1159 :     const int x, const int y,
1160 :     const uint32_t MotionFlags,
1161 :     const MBParam * const pParam,
1162 :     MACROBLOCK * const pMB,
1163 :     const MACROBLOCK * const pMBs,
1164 : Isibaar 539 const int block,
1165 :     SearchData * const Data)
1166 : chl 345 {
1167 : syskin 704 int i = 0;
1168 : Isibaar 539 Data->iMinSAD = OldData->iMinSAD + 1 + block;
1169 :     Data->currentMV = OldData->currentMV + 1 + block;
1170 : Isibaar 580 Data->currentQMV = OldData->currentQMV + 1 + block;
1171 : chl 345
1172 : syskin 601 if(pParam->m_quarterpel) {
1173 : syskin 704 Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1174 :     if (block != 0) i = d_mv_bits( Data->currentQMV->x - Data->predMV.x,
1175 : syskin 712 Data->currentQMV->y - Data->predMV.y, Data->iFcode, 0, 0);
1176 :    
1177 : Isibaar 606 } else {
1178 : syskin 704 Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1179 :     if (block != 0) {
1180 : syskin 712 if (block != 0) i = d_mv_bits( Data->currentMV->x - Data->predMV.x,
1181 :     Data->currentMV->y - Data->predMV.y, Data->iFcode, 0, Data->rrv);
1182 : syskin 704 }
1183 : Isibaar 606 }
1184 : chl 345
1185 : syskin 704 *(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100;
1186 :    
1187 : chl 530 if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) {
1188 : syskin 704 if (Data->rrv) i = 2; else i = 1;
1189 : chl 345
1190 : syskin 704 Data->Ref = OldData->Ref + i*8 * ((block&1) + pParam->edged_width*(block>>1));
1191 :     Data->RefH = OldData->RefH + i*8 * ((block&1) + pParam->edged_width*(block>>1));
1192 :     Data->RefV = OldData->RefV + i*8 * ((block&1) + pParam->edged_width*(block>>1));
1193 :     Data->RefHV = OldData->RefHV + i*8 * ((block&1) + pParam->edged_width*(block>>1));
1194 : chl 345
1195 : syskin 704 Data->Cur = OldData->Cur + i*8 * ((block&1) + pParam->edged_width*(block>>1));
1196 : syskin 663 Data->qpel_precision = 0;
1197 : chl 345
1198 : syskin 712 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1199 :     pParam->width, pParam->height, OldData->iFcode - Data->qpel, 0, Data->rrv);
1200 :    
1201 :     if (Data->rrv) CheckCandidate = CheckCandidate16no4v;
1202 :     else CheckCandidate = CheckCandidate8;
1203 :    
1204 : chl 530 if (MotionFlags & PMV_EXTSEARCH8) {
1205 : Isibaar 580 int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1206 :    
1207 : chl 530 MainSearchFunc *MainSearchPtr;
1208 :     if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch;
1209 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch;
1210 :     else MainSearchPtr = DiamondSearch;
1211 : chl 345
1212 : Isibaar 580 (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255);
1213 : chl 345
1214 : syskin 601 if(*(Data->iMinSAD) < temp_sad) {
1215 : Isibaar 580 Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1216 :     Data->currentQMV->y = 2 * Data->currentMV->y;
1217 :     }
1218 :     }
1219 :    
1220 :     if (MotionFlags & PMV_HALFPELREFINE8) {
1221 :     int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1222 :    
1223 : syskin 663 SubpelRefine(Data); // perform halfpel refine of current best vector
1224 : Isibaar 580
1225 :     if(*(Data->iMinSAD) < temp_sad) { // we have found a better match
1226 :     Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1227 :     Data->currentQMV->y = 2 * Data->currentMV->y;
1228 :     }
1229 :     }
1230 :    
1231 : syskin 712 if(!Data->rrv && Data->qpel) {
1232 : syskin 601 if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) &&
1233 :     (MotionFlags & PMV_QUARTERPELREFINE8)) {
1234 : syskin 663 Data->qpel_precision = 1;
1235 : syskin 712 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1236 :     pParam->width, pParam->height, OldData->iFcode, 1, 0);
1237 : syskin 663 SubpelRefine(Data);
1238 : syskin 601 }
1239 : Isibaar 580 }
1240 : chl 345 }
1241 : chl 530
1242 : syskin 704 if (Data->rrv) {
1243 :     Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x);
1244 :     Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y);
1245 :     }
1246 :    
1247 : syskin 712 if(Data->qpel) {
1248 : syskin 663 pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x;
1249 :     pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y;
1250 :     pMB->qmvs[block] = *(Data->currentQMV);
1251 : syskin 704 } else {
1252 : Isibaar 580 pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x;
1253 :     pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y;
1254 :     }
1255 :    
1256 : Isibaar 539 pMB->mvs[block] = *(Data->currentMV);
1257 : syskin 601 pMB->sad8[block] = 4 * (*Data->iMinSAD);
1258 : chl 345 }
1259 :    
1260 : chl 530 /* B-frames code starts here */
1261 : chl 345
1262 : chl 530 static __inline VECTOR
1263 :     ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode)
1264 : Isibaar 3 {
1265 : chl 530 /* the stupidiest function ever */
1266 :     if (mode == MODE_FORWARD) return pMB->mvs[0];
1267 :     else return pMB->b_mvs[0];
1268 : Isibaar 3 }
1269 :    
1270 : chl 530 static void __inline
1271 :     PreparePredictionsBF(VECTOR * const pmv, const int x, const int y,
1272 :     const uint32_t iWcount,
1273 :     const MACROBLOCK * const pMB,
1274 :     const uint32_t mode_curr)
1275 : Isibaar 3 {
1276 :    
1277 : chl 530 // [0] is prediction
1278 :     pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y);
1279 : Isibaar 3
1280 : chl 530 pmv[1].x = pmv[1].y = 0; // [1] is zero
1281 : Isibaar 3
1282 : chl 530 pmv[2] = ChoosePred(pMB, mode_curr);
1283 :     pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y);
1284 : edgomez 195
1285 : chl 530 if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour
1286 :     pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr);
1287 : Isibaar 539 pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y);
1288 :     } else pmv[3].x = pmv[3].y = 0;
1289 : Isibaar 3
1290 : chl 530 if (y != 0) {
1291 :     pmv[4] = ChoosePred(pMB-iWcount, mode_curr);
1292 :     pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y);
1293 :     } else pmv[4].x = pmv[4].y = 0;
1294 : Isibaar 3
1295 : chl 530 if (x != 0) {
1296 :     pmv[5] = ChoosePred(pMB-1, mode_curr);
1297 :     pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y);
1298 :     } else pmv[5].x = pmv[5].y = 0;
1299 : Isibaar 3
1300 : chl 530 if ((x != 0)&&(y != 0)) {
1301 :     pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr);
1302 :     pmv[6].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y);
1303 :     } else pmv[6].x = pmv[6].y = 0;
1304 : edgomez 195
1305 : chl 530 // more?
1306 :     }
1307 : chl 181
1308 : edgomez 170
1309 : chl 530 /* search backward or forward, for b-frames */
1310 :     static void
1311 :     SearchBF( const uint8_t * const pRef,
1312 :     const uint8_t * const pRefH,
1313 :     const uint8_t * const pRefV,
1314 :     const uint8_t * const pRefHV,
1315 :     const IMAGE * const pCur,
1316 :     const int x, const int y,
1317 :     const uint32_t MotionFlags,
1318 :     const uint32_t iFcode,
1319 :     const MBParam * const pParam,
1320 :     MACROBLOCK * const pMB,
1321 :     const VECTOR * const predMV,
1322 :     int32_t * const best_sad,
1323 : h 545 const int32_t mode_current,
1324 :     SearchData * const Data)
1325 : chl 530 {
1326 : Isibaar 3
1327 : chl 530 const int32_t iEdgedWidth = pParam->edged_width;
1328 :    
1329 :     int i, iDirection, mask;
1330 : h 545 VECTOR pmv[7];
1331 : chl 530 MainSearchFunc *MainSearchPtr;
1332 : h 545 *Data->iMinSAD = MV_MAX_ERROR;
1333 :     Data->iFcode = iFcode;
1334 : syskin 663 Data->qpel_precision = 0;
1335 : Isibaar 3
1336 : h 545 Data->Ref = pRef + (x + y * iEdgedWidth) * 16;
1337 :     Data->RefH = pRefH + (x + y * iEdgedWidth) * 16;
1338 :     Data->RefV = pRefV + (x + y * iEdgedWidth) * 16;
1339 :     Data->RefHV = pRefHV + (x + y * iEdgedWidth) * 16;
1340 : Isibaar 3
1341 : h 545 Data->predMV = *predMV;
1342 : edgomez 195
1343 : h 545 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1344 : syskin 712 pParam->width, pParam->height, iFcode - Data->qpel, 0, 0);
1345 : Isibaar 3
1346 : h 545 pmv[0] = Data->predMV;
1347 : syskin 658 if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; }
1348 : h 545 PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current);
1349 : edgomez 195
1350 : h 545 Data->currentMV->x = Data->currentMV->y = 0;
1351 : chl 530 CheckCandidate = CheckCandidate16no4v;
1352 : chl 181
1353 : chl 530 // main loop. checking all predictions
1354 :     for (i = 0; i < 8; i++) {
1355 :     if (!(mask = make_mask(pmv, i)) ) continue;
1356 : h 545 CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1357 : edgomez 78 }
1358 : Isibaar 3
1359 : chl 530 if (MotionFlags & PMV_USESQUARES16)
1360 :     MainSearchPtr = SquareSearch;
1361 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND16)
1362 :     MainSearchPtr = AdvDiamondSearch;
1363 :     else MainSearchPtr = DiamondSearch;
1364 : chl 96
1365 : h 545 (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255);
1366 : chl 169
1367 : syskin 663 SubpelRefine(Data);
1368 : syskin 658
1369 :     if (Data->qpel) {
1370 :     Data->currentQMV->x = 2*Data->currentMV->x;
1371 :     Data->currentQMV->y = 2*Data->currentMV->y;
1372 : syskin 663 Data->qpel_precision = 1;
1373 : syskin 712 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1374 :     pParam->width, pParam->height, iFcode, 1, 0);
1375 : syskin 663 SubpelRefine(Data);
1376 : syskin 658 }
1377 : chl 169
1378 : chl 530 // three bits are needed to code backward mode. four for forward
1379 :     // we treat the bits just like they were vector's
1380 : syskin 628 if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16;
1381 :     else *Data->iMinSAD += 3 * Data->lambda16;
1382 : chl 169
1383 : h 545 if (*Data->iMinSAD < *best_sad) {
1384 :     *best_sad = *Data->iMinSAD;
1385 : chl 530 pMB->mode = mode_current;
1386 : syskin 658 if (Data->qpel) {
1387 :     pMB->pmvs[0].x = Data->currentQMV->x - predMV->x;
1388 :     pMB->pmvs[0].y = Data->currentQMV->y - predMV->y;
1389 :     if (mode_current == MODE_FORWARD)
1390 :     pMB->qmvs[0] = *Data->currentQMV;
1391 :     else
1392 :     pMB->b_qmvs[0] = *Data->currentQMV;
1393 :     } else {
1394 :     pMB->pmvs[0].x = Data->currentMV->x - predMV->x;
1395 :     pMB->pmvs[0].y = Data->currentMV->y - predMV->y;
1396 :     }
1397 :     if (mode_current == MODE_FORWARD)
1398 :     pMB->mvs[0] = *(Data->currentMV+2) = *Data->currentMV;
1399 :     else
1400 :     pMB->b_mvs[0] = *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search
1401 :    
1402 : chl 169 }
1403 : chl 530
1404 :     }
1405 : chl 169
1406 : syskin 712 static void
1407 :     SkipDecisionB(const IMAGE * const pCur,
1408 :     const IMAGE * const f_Ref,
1409 :     const IMAGE * const b_Ref,
1410 :     MACROBLOCK * const pMB,
1411 :     const uint32_t quant,
1412 :     const uint32_t x, const uint32_t y,
1413 :     const SearchData * const Data)
1414 :     {
1415 :     int dx, dy, b_dx, b_dy;
1416 :     uint32_t sum;
1417 :     //this is not full chroma compensation, only it's fullpel approximation. should work though
1418 :     if (Data->qpel) {
1419 :     dy = Data->directmvF[0].y/2 + Data->directmvF[1].y/2 +
1420 :     Data->directmvF[2].y/2 + Data->directmvF[3].y/2;
1421 :    
1422 :     dx = Data->directmvF[0].x/2 + Data->directmvF[1].x/2 +
1423 :     Data->directmvF[2].x/2 + Data->directmvF[3].x/2;
1424 :    
1425 :     b_dy = Data->directmvB[0].y/2 + Data->directmvB[1].y/2 +
1426 :     Data->directmvB[2].y/2 + Data->directmvB[3].y/2;
1427 :    
1428 :     b_dx = Data->directmvB[0].x/2 + Data->directmvB[1].x/2 +
1429 :     Data->directmvB[2].x/2 + Data->directmvB[3].x/2;
1430 :    
1431 :     } else {
1432 :     dy = Data->directmvF[0].y + Data->directmvF[1].y +
1433 :     Data->directmvF[2].y + Data->directmvF[3].y;
1434 :    
1435 :     dx = Data->directmvF[0].x + Data->directmvF[1].x +
1436 :     Data->directmvF[2].x + Data->directmvF[3].x;
1437 :    
1438 :     b_dy = Data->directmvB[0].y + Data->directmvB[1].y +
1439 :     Data->directmvB[2].y + Data->directmvB[3].y;
1440 :    
1441 :     b_dx = Data->directmvB[0].x + Data->directmvB[1].x +
1442 :     Data->directmvB[2].x + Data->directmvB[3].x;
1443 :     }
1444 :    
1445 :    
1446 :     dy = (dy >> 3) + roundtab_76[dy & 0xf];
1447 :     dx = (dx >> 3) + roundtab_76[dx & 0xf];
1448 :     b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf];
1449 :     b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf];
1450 :    
1451 :     sum = sad8bi(pCur->u + 8*x + 8*y*(Data->iEdgedWidth/2),
1452 :     f_Ref->u + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2,
1453 :     b_Ref->u + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2,
1454 :     Data->iEdgedWidth/2);
1455 :     sum += sad8bi(pCur->v + 8*x + 8*y*(Data->iEdgedWidth/2),
1456 :     f_Ref->v + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2,
1457 :     b_Ref->v + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2,
1458 :     Data->iEdgedWidth/2);
1459 :    
1460 :     if (sum < 2*MAX_CHROMA_SAD_FOR_SKIP * quant) pMB->mode = MODE_DIRECT_NONE_MV; //skipped
1461 :     }
1462 :    
1463 :    
1464 :    
1465 :     static __inline uint32_t
1466 : Isibaar 539 SearchDirect(const IMAGE * const f_Ref,
1467 : chl 530 const uint8_t * const f_RefH,
1468 :     const uint8_t * const f_RefV,
1469 :     const uint8_t * const f_RefHV,
1470 : Isibaar 539 const IMAGE * const b_Ref,
1471 : chl 530 const uint8_t * const b_RefH,
1472 :     const uint8_t * const b_RefV,
1473 :     const uint8_t * const b_RefHV,
1474 :     const IMAGE * const pCur,
1475 :     const int x, const int y,
1476 :     const uint32_t MotionFlags,
1477 :     const int32_t TRB, const int32_t TRD,
1478 :     const MBParam * const pParam,
1479 :     MACROBLOCK * const pMB,
1480 :     const MACROBLOCK * const b_mb,
1481 : h 545 int32_t * const best_sad,
1482 :     SearchData * const Data)
1483 : edgomez 195
1484 : chl 530 {
1485 : h 545 int32_t skip_sad;
1486 : chl 530 int k;
1487 : h 545
1488 : chl 530 MainSearchFunc *MainSearchPtr;
1489 : Isibaar 3
1490 : h 545 *Data->iMinSAD = 256*4096;
1491 : chl 140
1492 : h 545 Data->Ref = f_Ref->y + (x + Data->iEdgedWidth*y) * 16;
1493 :     Data->RefH = f_RefH + (x + Data->iEdgedWidth*y) * 16;
1494 :     Data->RefV = f_RefV + (x + Data->iEdgedWidth*y) * 16;
1495 :     Data->RefHV = f_RefHV + (x + Data->iEdgedWidth*y) * 16;
1496 :     Data->bRef = b_Ref->y + (x + Data->iEdgedWidth*y) * 16;
1497 :     Data->bRefH = b_RefH + (x + Data->iEdgedWidth*y) * 16;
1498 :     Data->bRefV = b_RefV + (x + Data->iEdgedWidth*y) * 16;
1499 :     Data->bRefHV = b_RefHV + (x + Data->iEdgedWidth*y) * 16;
1500 : chl 140
1501 : h 545 Data->max_dx = 2 * pParam->width - 2 * (x) * 16;
1502 :     Data->max_dy = 2 * pParam->height - 2 * (y) * 16;
1503 :     Data->min_dx = -(2 * 16 + 2 * (x) * 16);
1504 :     Data->min_dy = -(2 * 16 + 2 * (y) * 16);
1505 : syskin 658 if (Data->qpel) { //we measure in qpixels
1506 :     Data->max_dx *= 2;
1507 :     Data->max_dy *= 2;
1508 :     Data->min_dx *= 2;
1509 :     Data->min_dy *= 2;
1510 :     Data->referencemv = b_mb->qmvs;
1511 :     } else Data->referencemv = b_mb->mvs;
1512 : syskin 672 Data->qpel_precision = 0; // it's a trick. it's 1 not 0, but we need 0 here
1513 : Isibaar 3
1514 : chl 530 for (k = 0; k < 4; k++) {
1515 : h 545 pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD);
1516 :     pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD;
1517 :     pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD);
1518 :     pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD;
1519 : Isibaar 3
1520 : h 545 if ( ( pMB->b_mvs[k].x > Data->max_dx ) || ( pMB->b_mvs[k].x < Data->min_dx )
1521 :     || ( pMB->b_mvs[k].y > Data->max_dy ) || ( pMB->b_mvs[k].y < Data->min_dy )) {
1522 : Isibaar 3
1523 : Isibaar 539 *best_sad = 256*4096; // in that case, we won't use direct mode
1524 :     pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV"
1525 :     pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0;
1526 : syskin 712 return 256*4096;
1527 : Isibaar 539 }
1528 :     if (b_mb->mode != MODE_INTER4V) {
1529 :     pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0];
1530 :     pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0];
1531 : h 545 Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0];
1532 :     Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0];
1533 : Isibaar 539 break;
1534 :     }
1535 :     }
1536 : Isibaar 3
1537 : syskin 663
1538 :     if (b_mb->mode == MODE_INTER4V) CheckCandidate = CheckCandidateDirect;
1539 :     else CheckCandidate = CheckCandidateDirectno4v;
1540 : syskin 658
1541 : h 545 (*CheckCandidate)(0, 0, 255, &k, Data);
1542 : Isibaar 3
1543 : syskin 712 // initial (fast) skip decision
1544 :     if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH*2) {
1545 :     SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data->chroma, Data); //possible skip - checking chroma
1546 :     if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip.
1547 : Isibaar 539 }
1548 :    
1549 : h 545 skip_sad = *Data->iMinSAD;
1550 : edgomez 195
1551 : chl 530 // DIRECT MODE DELTA VECTOR SEARCH.
1552 :     // This has to be made more effective, but at the moment I'm happy it's running at all
1553 : Isibaar 3
1554 : chl 530 if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1555 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1556 :     else MainSearchPtr = DiamondSearch;
1557 : edgomez 195
1558 : h 545 (*MainSearchPtr)(0, 0, Data, 255);
1559 : Isibaar 3
1560 : syskin 663 SubpelRefine(Data);
1561 : Isibaar 3
1562 : h 545 *best_sad = *Data->iMinSAD;
1563 : Isibaar 3
1564 : syskin 672 if (b_mb->mode == MODE_INTER4V)
1565 : chl 530 pMB->mode = MODE_DIRECT;
1566 : syskin 672 else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation
1567 : edgomez 195
1568 : h 545 pMB->pmvs[3] = *Data->currentMV;
1569 : edgomez 195
1570 : chl 530 for (k = 0; k < 4; k++) {
1571 : h 545 pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x;
1572 : syskin 658 pMB->b_mvs[k].x = ( (Data->currentMV->x == 0)
1573 : h 545 ? Data->directmvB[k].x
1574 : syskin 658 :pMB->mvs[k].x - Data->referencemv[k].x);
1575 : h 545 pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y);
1576 :     pMB->b_mvs[k].y = ((Data->currentMV->y == 0)
1577 :     ? Data->directmvB[k].y
1578 :     : pMB->mvs[k].y - Data->referencemv[k].y);
1579 : syskin 658 if (Data->qpel) {
1580 :     pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2;
1581 :     pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2;
1582 :     pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2;
1583 :     pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2;
1584 :     }
1585 :    
1586 : chl 530 if (b_mb->mode != MODE_INTER4V) {
1587 :     pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0];
1588 :     pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0];
1589 : syskin 658 pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0];
1590 :     pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0];
1591 : chl 530 break;
1592 :     }
1593 :     }
1594 : h 545 return skip_sad;
1595 : Isibaar 3 }
1596 : chl 96
1597 : h 545
1598 : syskin 712 static void
1599 : chl 530 SearchInterpolate(const uint8_t * const f_Ref,
1600 :     const uint8_t * const f_RefH,
1601 :     const uint8_t * const f_RefV,
1602 :     const uint8_t * const f_RefHV,
1603 :     const uint8_t * const b_Ref,
1604 :     const uint8_t * const b_RefH,
1605 :     const uint8_t * const b_RefV,
1606 :     const uint8_t * const b_RefHV,
1607 :     const IMAGE * const pCur,
1608 :     const int x, const int y,
1609 :     const uint32_t fcode,
1610 :     const uint32_t bcode,
1611 :     const uint32_t MotionFlags,
1612 :     const MBParam * const pParam,
1613 :     const VECTOR * const f_predMV,
1614 :     const VECTOR * const b_predMV,
1615 :     MACROBLOCK * const pMB,
1616 : h 545 int32_t * const best_sad,
1617 :     SearchData * const fData)
1618 : chl 530
1619 : chl 96 {
1620 :    
1621 : chl 530 int iDirection, i, j;
1622 : h 545 SearchData bData;
1623 : chl 96
1624 : syskin 712 fData->qpel_precision = 0;
1625 :     memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data
1626 :     *fData->iMinSAD = 4096*256;
1627 :     bData.currentMV ++; bData.currentQMV ++;
1628 : h 545 fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode;
1629 : edgomez 195
1630 : syskin 712 i = (x + y * fData->iEdgedWidth) * 16;
1631 :     bData.bRef = fData->Ref = f_Ref + i;
1632 :     bData.bRefH = fData->RefH = f_RefH + i;
1633 :     bData.bRefV = fData->RefV = f_RefV + i;
1634 :     bData.bRefHV = fData->RefHV = f_RefHV + i;
1635 :     bData.Ref = fData->bRef = b_Ref + i;
1636 :     bData.RefH = fData->bRefH = b_RefH + i;
1637 :     bData.RefV = fData->bRefV = b_RefV + i;
1638 :     bData.RefHV = fData->bRefHV = b_RefHV + i;
1639 : chl 96
1640 : h 545 bData.bpredMV = fData->predMV = *f_predMV;
1641 :     fData->bpredMV = bData.predMV = *b_predMV;
1642 : syskin 661 fData->currentMV[0] = fData->currentMV[2];
1643 : chl 96
1644 : syskin 712 get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode - fData->qpel, 0, 0);
1645 :     get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode - fData->qpel, 0, 0);
1646 :    
1647 : h 545 if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx;
1648 : syskin 675 if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx;
1649 :     if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy;
1650 :     if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy;
1651 : Isibaar 539
1652 : h 545 if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx;
1653 : syskin 675 if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx;
1654 :     if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy;
1655 :     if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy;
1656 : Isibaar 539
1657 : h 545 CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData);
1658 : chl 96
1659 : chl 530 //diamond. I wish we could use normal mainsearch functions (square, advdiamond)
1660 : chl 96
1661 : chl 530 do {
1662 :     iDirection = 255;
1663 :     // forward MV moves
1664 : h 545 i = fData->currentMV[0].x; j = fData->currentMV[0].y;
1665 : chl 96
1666 : h 545 CheckCandidateInt(i + 1, j, 0, &iDirection, fData);
1667 :     CheckCandidateInt(i, j + 1, 0, &iDirection, fData);
1668 :     CheckCandidateInt(i - 1, j, 0, &iDirection, fData);
1669 :     CheckCandidateInt(i, j - 1, 0, &iDirection, fData);
1670 : chl 96
1671 : chl 530 // backward MV moves
1672 : h 545 i = fData->currentMV[1].x; j = fData->currentMV[1].y;
1673 :     fData->currentMV[2] = fData->currentMV[0];
1674 : Isibaar 539 CheckCandidateInt(i + 1, j, 0, &iDirection, &bData);
1675 :     CheckCandidateInt(i, j + 1, 0, &iDirection, &bData);
1676 :     CheckCandidateInt(i - 1, j, 0, &iDirection, &bData);
1677 :     CheckCandidateInt(i, j - 1, 0, &iDirection, &bData);
1678 : chl 96
1679 : chl 530 } while (!(iDirection));
1680 : edgomez 195
1681 : syskin 658 if (fData->qpel) {
1682 : syskin 675 CheckCandidate = CheckCandidateInt;
1683 : syskin 663 fData->qpel_precision = bData.qpel_precision = 1;
1684 : syskin 712 get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0);
1685 :     get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0);
1686 : syskin 658 fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x;
1687 :     fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y;
1688 :     fData->currentQMV[1].x = 2 * fData->currentMV[1].x;
1689 :     fData->currentQMV[1].y = 2 * fData->currentMV[1].y;
1690 : syskin 663 SubpelRefine(fData);
1691 : syskin 658 fData->currentQMV[2] = fData->currentQMV[0];
1692 : syskin 663 SubpelRefine(&bData);
1693 : syskin 658 }
1694 : syskin 676
1695 : syskin 712 *fData->iMinSAD += (2+2) * fData->lambda16; // two bits are needed to code interpolate mode.
1696 : syskin 658
1697 : h 545 if (*fData->iMinSAD < *best_sad) {
1698 :     *best_sad = *fData->iMinSAD;
1699 :     pMB->mvs[0] = fData->currentMV[0];
1700 :     pMB->b_mvs[0] = fData->currentMV[1];
1701 : chl 530 pMB->mode = MODE_INTERPOLATE;
1702 : syskin 658 if (fData->qpel) {
1703 :     pMB->qmvs[0] = fData->currentQMV[0];
1704 :     pMB->b_qmvs[0] = fData->currentQMV[1];
1705 :     pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x;
1706 :     pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y;
1707 :     pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x;
1708 :     pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y;
1709 :     } else {
1710 :     pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x;
1711 :     pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y;
1712 :     pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x;
1713 :     pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y;
1714 :     }
1715 : chl 530 }
1716 :     }
1717 : chl 96
1718 : chl 530 void
1719 :     MotionEstimationBVOP(MBParam * const pParam,
1720 :     FRAMEINFO * const frame,
1721 :     const int32_t time_bp,
1722 :     const int32_t time_pp,
1723 :     // forward (past) reference
1724 :     const MACROBLOCK * const f_mbs,
1725 :     const IMAGE * const f_ref,
1726 :     const IMAGE * const f_refH,
1727 :     const IMAGE * const f_refV,
1728 :     const IMAGE * const f_refHV,
1729 :     // backward (future) reference
1730 : syskin 644 const FRAMEINFO * const b_reference,
1731 : chl 530 const IMAGE * const b_ref,
1732 :     const IMAGE * const b_refH,
1733 :     const IMAGE * const b_refV,
1734 :     const IMAGE * const b_refHV)
1735 :     {
1736 :     uint32_t i, j;
1737 : syskin 712 int32_t best_sad;
1738 :     uint32_t skip_sad;
1739 : chl 530 int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0;
1740 :     static const VECTOR zeroMV={0,0};
1741 : syskin 644 const MACROBLOCK * const b_mbs = b_reference->mbs;
1742 : chl 96
1743 : chl 530 VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/
1744 : chl 96
1745 : chl 530 const int32_t TRB = time_pp - time_bp;
1746 :     const int32_t TRD = time_pp;
1747 : syskin 658 uint8_t * qimage;
1748 : chl 96
1749 : h 545 // some pre-inintialized data for the rest of the search
1750 :    
1751 :     SearchData Data;
1752 :     int32_t iMinSAD;
1753 :     VECTOR currentMV[3];
1754 : syskin 658 VECTOR currentQMV[3];
1755 : syskin 704 memset(&Data, 0, sizeof(SearchData));
1756 : h 545 Data.iEdgedWidth = pParam->edged_width;
1757 : syskin 658 Data.currentMV = currentMV; Data.currentQMV = currentQMV;
1758 : h 545 Data.iMinSAD = &iMinSAD;
1759 : syskin 712 Data.lambda16 = lambda_vec16[frame->quant];
1760 :     Data.chroma = frame->quant;
1761 : syskin 658 Data.qpel = pParam->m_quarterpel;
1762 : syskin 663 Data.rounding = 0;
1763 : h 545
1764 : syskin 658 if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL)
1765 :     return; // allocate some mem for qpel interpolated blocks
1766 :     // somehow this is dirty since I think we shouldn't use malloc outside
1767 :     // encoder_create() - so please fix me!
1768 :     Data.RefQ = qimage;
1769 :    
1770 : chl 530 // note: i==horizontal, j==vertical
1771 :     for (j = 0; j < pParam->mb_height; j++) {
1772 : chl 96
1773 : chl 530 f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */
1774 : edgomez 195
1775 : chl 530 for (i = 0; i < pParam->mb_width; i++) {
1776 :     MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width;
1777 :     const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width;
1778 : edgomez 195
1779 : chl 619 /* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */
1780 : syskin 644 if (b_reference->coding_type != S_VOP)
1781 :     if (b_mb->mode == MODE_NOT_CODED) {
1782 :     pMB->mode = MODE_NOT_CODED;
1783 :     continue;
1784 :     }
1785 : chl 96
1786 : h 545 Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16;
1787 : syskin 601 pMB->quant = frame->quant;
1788 : syskin 628
1789 : chl 530 /* direct search comes first, because it (1) checks for SKIP-mode
1790 :     and (2) sets very good predictions for forward and backward search */
1791 : Isibaar 539 skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y,
1792 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
1793 : chl 530 &frame->image,
1794 :     i, j,
1795 :     frame->motion_flags,
1796 :     TRB, TRD,
1797 :     pParam,
1798 :     pMB, b_mb,
1799 : h 545 &best_sad,
1800 :     &Data);
1801 : chl 96
1802 : Isibaar 539 if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; }
1803 : syskin 658
1804 : chl 530 // forward search
1805 :     SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y,
1806 :     &frame->image, i, j,
1807 :     frame->motion_flags,
1808 : h 545 frame->fcode, pParam,
1809 : chl 530 pMB, &f_predMV, &best_sad,
1810 : h 545 MODE_FORWARD, &Data);
1811 : chl 96
1812 : chl 530 // backward search
1813 :     SearchBF(b_ref->y, b_refH->y, b_refV->y, b_refHV->y,
1814 :     &frame->image, i, j,
1815 :     frame->motion_flags,
1816 : h 545 frame->bcode, pParam,
1817 : chl 530 pMB, &b_predMV, &best_sad,
1818 : h 545 MODE_BACKWARD, &Data);
1819 : chl 96
1820 : chl 530 // interpolate search comes last, because it uses data from forward and backward as prediction
1821 : syskin 662
1822 : chl 530 SearchInterpolate(f_ref->y, f_refH->y, f_refV->y, f_refHV->y,
1823 :     b_ref->y, b_refH->y, b_refV->y, b_refHV->y,
1824 :     &frame->image,
1825 :     i, j,
1826 :     frame->fcode, frame->bcode,
1827 :     frame->motion_flags,
1828 : h 545 pParam,
1829 : chl 530 &f_predMV, &b_predMV,
1830 : h 545 pMB, &best_sad,
1831 :     &Data);
1832 : syskin 662
1833 : syskin 712 // final skip decision
1834 :     if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP*2)
1835 :     && ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) )
1836 :     SkipDecisionB(&frame->image, f_ref, b_ref, pMB,frame->quant, i, j, &Data);
1837 :    
1838 : chl 530 switch (pMB->mode) {
1839 :     case MODE_FORWARD:
1840 :     f_count++;
1841 : syskin 712 if (Data.qpel) f_predMV = pMB->qmvs[0];
1842 : syskin 658 else f_predMV = pMB->mvs[0];
1843 : chl 530 break;
1844 :     case MODE_BACKWARD:
1845 :     b_count++;
1846 : syskin 712 if (Data.qpel) b_predMV = pMB->b_qmvs[0];
1847 : syskin 658 else b_predMV = pMB->b_mvs[0];
1848 : chl 530 break;
1849 :     case MODE_INTERPOLATE:
1850 :     i_count++;
1851 : syskin 712 if (Data.qpel) {
1852 : syskin 658 f_predMV = pMB->qmvs[0];
1853 :     b_predMV = pMB->b_qmvs[0];
1854 :     } else {
1855 :     f_predMV = pMB->mvs[0];
1856 :     b_predMV = pMB->b_mvs[0];
1857 :     }
1858 : chl 530 break;
1859 :     case MODE_DIRECT:
1860 :     case MODE_DIRECT_NO4V:
1861 :     d_count++;
1862 :     default:
1863 :     break;
1864 : chl 96 }
1865 :     }
1866 :     }
1867 : syskin 658 free(qimage);
1868 : chl 96 }
1869 :    
1870 : Isibaar 539 static __inline int
1871 :     MEanalyzeMB ( const uint8_t * const pRef,
1872 :     const uint8_t * const pCur,
1873 :     const int x,
1874 :     const int y,
1875 :     const MBParam * const pParam,
1876 :     const MACROBLOCK * const pMBs,
1877 : h 545 MACROBLOCK * const pMB,
1878 :     SearchData * const Data)
1879 : Isibaar 539 {
1880 : chl 317
1881 : Isibaar 606 int i = 255, mask;
1882 : h 545 VECTOR pmv[3];
1883 : syskin 640 *(Data->iMinSAD) = MV_MAX_ERROR;
1884 : suxen_drol 118
1885 : syskin 640 //median is only used as prediction. it doesn't have to be real
1886 :     if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0;
1887 :     else
1888 :     if (x == 1) //left macroblock does not have any vector now
1889 :     Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median
1890 :     else if (y == 1) // top macroblock don't have it's vector
1891 :     Data->predMV = (pMB - 1)->mvs[0]; // left instead of median
1892 :     else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median
1893 :    
1894 : h 545 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1895 : syskin 712 pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, 0);
1896 : chl 326
1897 : h 545 Data->Cur = pCur + (x + y * pParam->edged_width) * 16;
1898 :     Data->Ref = pRef + (x + y * pParam->edged_width) * 16;
1899 : Isibaar 539
1900 :     pmv[1].x = EVEN(pMB->mvs[0].x);
1901 :     pmv[1].y = EVEN(pMB->mvs[0].y);
1902 : syskin 640 pmv[2].x = EVEN(Data->predMV.x);
1903 :     pmv[2].y = EVEN(Data->predMV.y);
1904 :     pmv[0].x = pmv[0].y = 0;
1905 : chl 326
1906 : syskin 712 CheckCandidate16no4vI(0, 0, 255, &i, Data);
1907 : syskin 640
1908 :     //early skip for 0,0
1909 :     if (*Data->iMinSAD < MAX_SAD00_FOR_SKIP * 4) {
1910 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
1911 :     pMB->mode = MODE_NOT_CODED;
1912 :     return 0;
1913 :     }
1914 :    
1915 : Isibaar 539 if (!(mask = make_mask(pmv, 1)))
1916 : syskin 712 CheckCandidate16no4vI(pmv[1].x, pmv[1].y, mask, &i, Data);
1917 : Isibaar 539 if (!(mask = make_mask(pmv, 2)))
1918 : syskin 712 CheckCandidate16no4vI(pmv[2].x, pmv[2].y, mask, &i, Data);
1919 : Isibaar 539
1920 : syskin 712 if (*Data->iMinSAD > MAX_SAD00_FOR_SKIP * 6) // diamond only if needed
1921 : syskin 640 DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i);
1922 : Isibaar 539
1923 : syskin 640 pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
1924 : syskin 628 pMB->mode = MODE_INTER;
1925 : h 545 return *(Data->iMinSAD);
1926 : Isibaar 539 }
1927 :    
1928 :     #define INTRA_THRESH 1350
1929 : syskin 672 #define INTER_THRESH 1200
1930 : Isibaar 539
1931 : syskin 644
1932 : Isibaar 539 int
1933 :     MEanalysis( const IMAGE * const pRef,
1934 : syskin 644 FRAMEINFO * const Current,
1935 : Isibaar 539 MBParam * const pParam,
1936 : syskin 644 int maxIntra, //maximum number if non-I frames
1937 :     int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame
1938 :     int bCount) // number if B frames in a row
1939 : Isibaar 539 {
1940 :     uint32_t x, y, intra = 0;
1941 :     int sSAD = 0;
1942 : syskin 644 MACROBLOCK * const pMBs = Current->mbs;
1943 :     const IMAGE * const pCurrent = &Current->image;
1944 :     int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH;
1945 : h 545
1946 :     VECTOR currentMV;
1947 :     int32_t iMinSAD;
1948 :     SearchData Data;
1949 :     Data.iEdgedWidth = pParam->edged_width;
1950 :     Data.currentMV = &currentMV;
1951 :     Data.iMinSAD = &iMinSAD;
1952 : syskin 644 Data.iFcode = Current->fcode;
1953 : syskin 640 CheckCandidate = CheckCandidate16no4vI;
1954 : h 545
1955 : syskin 658 if (intraCount < 10) // we're right after an I frame
1956 :     IntraThresh += 4 * (intraCount - 10) * (intraCount - 10);
1957 : syskin 644 else
1958 :     if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec
1959 :     IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra;
1960 :    
1961 :    
1962 : syskin 672 InterThresh += 400 * (1 - bCount);
1963 : syskin 644 if (InterThresh < 200) InterThresh = 200;
1964 :    
1965 : Isibaar 539 if (sadInit) (*sadInit) ();
1966 :    
1967 : syskin 704 for (y = 1; y < pParam->mb_height-1; y++) {
1968 :     for (x = 1; x < pParam->mb_width-1; x++) {
1969 : Isibaar 539 int sad, dev;
1970 :     MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width];
1971 :    
1972 :     sad = MEanalyzeMB(pRef->y, pCurrent->y, x, y,
1973 : h 545 pParam, pMBs, pMB, &Data);
1974 : Isibaar 539
1975 : syskin 644 if (sad > IntraThresh) {
1976 : syskin 601 dev = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16,
1977 :     pParam->edged_width);
1978 : syskin 644 if (dev + IntraThresh < sad) {
1979 :     pMB->mode = MODE_INTRA;
1980 : suxen_drol 709 if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return I_VOP;
1981 : Isibaar 539 }
1982 : syskin 644 }
1983 : syskin 601 sSAD += sad;
1984 : suxen_drol 118 }
1985 :     }
1986 : syskin 704 sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2);
1987 : suxen_drol 709 if (sSAD > InterThresh ) return P_VOP;
1988 : Isibaar 539 emms();
1989 : suxen_drol 709 return B_VOP;
1990 : syskin 712
1991 : suxen_drol 118 }
1992 : syskin 576
1993 : syskin 628 static void
1994 :     CheckGMC(int x, int y, const int dir, int * iDirection,
1995 :     const MACROBLOCK * const pMBs, uint32_t * bestcount, VECTOR * GMC,
1996 :     const MBParam * const pParam)
1997 :     {
1998 :     uint32_t mx, my, a, count = 0;
1999 : chl 619
2000 : syskin 628 for (my = 1; my < pParam->mb_height-1; my++)
2001 :     for (mx = 1; mx < pParam->mb_width-1; mx++) {
2002 :     VECTOR mv;
2003 :     const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width];
2004 :     if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) continue;
2005 :     mv = pMB->mvs[0];
2006 :     a = ABS(mv.x - x) + ABS(mv.y - y);
2007 :     if (a < 6) count += 6 - a;
2008 :     }
2009 : chl 619
2010 : syskin 628 if (count > *bestcount) {
2011 :     *bestcount = count;
2012 :     *iDirection = dir;
2013 :     GMC->x = x; GMC->y = y;
2014 :     }
2015 :     }
2016 :    
2017 :    
2018 : syskin 712 static __inline VECTOR
2019 : chl 619 GlobalMotionEst(const MACROBLOCK * const pMBs, const MBParam * const pParam, const uint32_t iFcode)
2020 :     {
2021 :    
2022 : syskin 628 uint32_t count, bestcount = 0;
2023 : chl 619 int x, y;
2024 : syskin 628 VECTOR gmc = {0,0};
2025 : chl 619 int step, min_x, max_x, min_y, max_y;
2026 :     uint32_t mx, my;
2027 : syskin 628 int iDirection, bDirection;
2028 : chl 619
2029 :     min_x = min_y = -32<<iFcode;
2030 :     max_x = max_y = 32<<iFcode;
2031 :    
2032 : syskin 628 //step1: let's find a rough camera panning
2033 : chl 619 for (step = 32; step >= 2; step /= 2) {
2034 :     bestcount = 0;
2035 :     for (y = min_y; y <= max_y; y += step)
2036 :     for (x = min_x ; x <= max_x; x += step) {
2037 :     count = 0;
2038 :     //for all macroblocks
2039 :     for (my = 1; my < pParam->mb_height-1; my++)
2040 :     for (mx = 1; mx < pParam->mb_width-1; mx++) {
2041 :     const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width];
2042 :     VECTOR mv;
2043 :    
2044 :     if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED)
2045 :     continue;
2046 :    
2047 :     mv = pMB->mvs[0];
2048 :     if ( ABS(mv.x - x) <= step && ABS(mv.y - y) <= step ) /* GMC translation is always halfpel-res */
2049 :     count++;
2050 :     }
2051 :     if (count >= bestcount) { bestcount = count; gmc.x = x; gmc.y = y; }
2052 :     }
2053 :     min_x = gmc.x - step;
2054 :     max_x = gmc.x + step;
2055 :     min_y = gmc.y - step;
2056 :     max_y = gmc.y + step;
2057 :    
2058 :     }
2059 : syskin 628
2060 :     if (bestcount < (pParam->mb_height-2)*(pParam->mb_width-2)/10)
2061 :     gmc.x = gmc.y = 0; //no camara pan, no GMC
2062 :    
2063 :     // step2: let's refine camera panning using gradiend-descent approach.
2064 :     // TODO: more warping points may be evaluated here (like in interpolate mode search - two vectors in one diamond)
2065 :     bestcount = 0;
2066 :     CheckGMC(gmc.x, gmc.y, 255, &iDirection, pMBs, &bestcount, &gmc, pParam);
2067 :     do {
2068 :     x = gmc.x; y = gmc.y;
2069 :     bDirection = iDirection; iDirection = 0;
2070 :     if (bDirection & 1) CheckGMC(x - 1, y, 1+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam);
2071 :     if (bDirection & 2) CheckGMC(x + 1, y, 2+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam);
2072 :     if (bDirection & 4) CheckGMC(x, y - 1, 1+2+4, &iDirection, pMBs, &bestcount, &gmc, pParam);
2073 :     if (bDirection & 8) CheckGMC(x, y + 1, 1+2+8, &iDirection, pMBs, &bestcount, &gmc, pParam);
2074 :    
2075 :     } while (iDirection);
2076 :    
2077 : chl 619 if (pParam->m_quarterpel) {
2078 :     gmc.x *= 2;
2079 :     gmc.y *= 2; /* we store the halfpel value as pseudo-qpel to make comparison easier */
2080 :     }
2081 :    
2082 :     return gmc;
2083 :     }

No admin address has been configured
ViewVC Help
Powered by ViewVC 1.0.4